Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
AdaTS: Learning Adaptive Time Series Representations via Dynamic Soft Contrasts
Authors: Denizhan Kara, Tomoyoshi Kimura, Jinyang Li, Bowen He, Yizhuo Chen, Yigong Hu, Hongjue Zhao, Shengzhong Liu, Tarek Abdelzaher
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We extensively evaluate Ada TS across six TS datasets, demonstrating its ability to augment existing methods and enhance performance. Specifically, Ada TS improves average CL accuracy by 7.3% (up to 13.7%), enhances robustness to dynamic variations, and achieves superior performance at low label rates. Our results show that Ada TS effectively incorporates underlying physical and temporal correlations and adapts to the varying sequence dynamics common in real-world TS applications. |
| Researcher Affiliation | Academia | University of Illinois Urbana-Champaign Shanghai Jiao Tong University EMAIL EMAIL |
| Pseudocode | Yes | Algorithm 1 TIMEFREQUENCYCOHERENCE Algorithm 2 ORDINAL CONSISTENCY LOSS Algorithm 3 DYNAMIC TEMPORAL CONTRASTIVE LOSS |
| Open Source Code | Yes | Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? Answer: [Yes] Justification: The code is available at https://github.com/denizhankara/Ada TS. |
| Open Datasets | Yes | We extensively evaluate Ada TS across two Io T application domains using four datasets: Human Activity Recognition (PAMAP2 [48], RWHAR [53]) and Moving Object Detection (MOD [37], ACIDS [62]). We additionally evaluate on standard time series classification benchmarks: UCR Archive (univariate) [12] and UEA Archive (multivariate) [3]. |
| Dataset Splits | Yes | We partition each dataset into training, validation, and test sets using an 8:1:1 ratio. Detailed dataset descriptions and configurations are provided in Appendix A. ... We employ an 8:1:1 random split at the run level for training, validation, and testing respectively. ... We partition samples randomly at the sample level with an 8:1:1 ratio for training, validation, and testing. |
| Hardware Specification | Yes | Evaluations are performed on NVIDIA RTX 6000 Ada GPUs with 48GB of memory. |
| Software Dependencies | Yes | We develop the code from the open-source implementations of foundational models and techniques [7, 37, 39, 34] using Py Torch 2.0.1. |
| Experiment Setup | Yes | This section details the training strategies, hyperparameter settings, and implementation specifics for pretraining and fine-tuning models with Ada TS. The primary configurations are summarized in Table 13, and SW-T model-specific parameters are in Table 14. ... Table 13: General Training Configurations for Ada TS. Parameter Pretraining Fine-tuning Optimizer Adam W Adam Weight Decay 0.05 0.05 Max Learning Rate 1e-4 (Default) 1e-2 (Default) Min Learning Rate (Pretrain) 1e-7 N/A Learning Rate Scheduler Cosine Annealing Step Decay LR Decay Factor (Fine-tuning) N/A 0.2 LR Decay Period (Fine-tuning) N/A 50 epochs Warmup Epochs (Pretrain) 10 N/A Epochs MOD, ACIDS: 2500 RWHAR, PAMAP2: 1000 200 Batch Size 256 (64 sequences) 128 Sequence Length (Tseq) 4 N/A Temperature (τ) 0.1 (Default for CL) N/A |